FLARE-BO: Fused Luminance and Adaptive Retinex Enhancement via Bayesian Optimisation for Low-Light Robotic Vision

arXiv:2604.22093 · cs.CV, eess.IV · Submitted 2026-04-23 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "FLARE-BO: Fused Luminance and Adaptive Retinex Enhancement via Bayesian Optimisation for Low-Light Robotic Vision".

Jane: Here's the gist of the paper. The gist The proposed FLARE-BO framework extends Bayesian optimisation to jointly optimise eight parameters spanning across gamma correction, LIME-style illumination normalisation, chrominance denoising, bilateral filtering, NLM denoising, GreyWorld automatic white balance, and adaptive post smoothing for low-light image enhancement.

Tom: This method achieves marked improvements over existing methods when benchmarked using the Low Light paired dataset (LOL) How it works FLARE-BO processes each low-light image I through a structured eight-stage pipeline whose parameters θ = (α, β, γ, h, σs, λ, d, hc) are selected per image by Bayesian Optimisation The key contributions include a "structured eight-stage enhancement pipeline incorporating LIME style guided illumination normalisation" and other specific components controlled by dedicated optimised parameters.

Jane: The eight stages of the pipeline are as follows: 1.

Tom: Stage 1 involves LIME-Style Illumination Normalisation, where when λ > 0, a Retinex-inspired illumination correction is applied first to estimate the per-pixel illumination map and then normalise the image.

Title and authors: Jane: So, what exactly is this paper proposing with FLARE-BO? It’s titled "FLARE-BO: Fused Luminance and Adaptive Retinex Enhancement via Bayesian Optimisation for Low-Light Robotic Vision".

Tom: It sounds like they are fusing luminance and adaptive retinex enhancement through the use of Bayesian optimisation. That means they aren't using one fixed algorithm, but a system that learns the best settings for each specific low-light picture it sees.

Lu: The authors are Nathan Shankar, Pawel Ladosz, and Hujun Yin. They are proposing this as a training-free approach to reliable visual perception under low illumination for autonomous robotic systems <ref:2604.22093#pg1>.

Meng: From what I see in the introduction, the main implication is that robots can perform navigation or inspection even when the lighting conditions change unexpectedly, because this system adapts dynamically.

Lalam: They achieved competitive enhancement without needing any learned model, which means we don't need massive datasets to train something like RetinexNet ten if we want good performance <ref:2604.22093#pg1>.

Tom: The paper then immediately points out the limitations of the earlier framework they built upon; it only had three parameters and relied on Non-Local Means denoising, which tends to over smooth edges under noisy conditions.

Jane: So, what’s the main contribution of FLARE-BO in terms of what it actually *does* differently from those earlier methods?

The paper's summary: Lu: FLARE-BO extends the previous work by jointly optimising eight parameters that span across gamma correction, LIME-style illumination normalisation, chrominance denoising, bilateral filtering, NLM denoising, GreyWorld automatic white balance, and adaptive post smoothing.

Tom: So it’s adding more complexity to the optimisation space than before—eight parameters instead of three—and incorporating specific steps like LIME style illumination normalisation and grey world white balance when those settings are used <ref:2604.22093#pg1>.

Meng: The summary emphasizes that this is a structured eight-stage enhancement pipeline where the parameters are selected per image by Bayesian optimisation. That sounds like it has a lot of computational overhead, which I need to think about for real-time deployment.

Jane: It seems the key takeaway here is that this framework tackles several problems at once—illumination, colour bias, noise suppression, and structural preservation—under one unified adaptive search process.

Lalam: The authors are showing that by jointly optimising these eight things, they can achieve marked improvements over existing methods when benchmarked using the Low Light paired dataset LOL <ref:2604.22093#pg1>.

Tom: So, if we translate that into plain language, it means this method handles a lot of image problems simultaneously by finding the right combination of eight adjustments for every single low-light picture.

The paper's improvements: Jane: The paper details several specific components that make up this extension. They introduce a "structured eight-stage enhancement pipeline incorporating LIME style guided illumination normalisation" as one of the key contributions.

Tom: That first stage specifically involves applying a Retinex-inspired illumination correction when the parameter lambda is greater than zero, and then normalising the image <ref:2604.22093#pg1>.

Lu: Then they have Gamma and Linear Contrast Adjustment using a formula like I' = clip alpha * I gamma lime * two hundred fifty-five + beta, zero two hundred fifty-five <ref:2604.22093#pg1>. The parameters alpha, gamma, and beta are all controlled by the optimiser.

Meng: And then they move into colour correction with Stage three involving Grey World Automatic White Balance correction to remove chromatic bias if the parameter hc is greater than zero <ref:2604.22093#pg1>. That sounds like it’s tackling both light level and colour issues together.

Lalam: Then Stage four applies LAB Space Chrominance Denoising when hc is greater than zero, using a median filter separately on the a star and b star chrominance channels <ref:2604.22093#pg3>.

Tom: And Stage five involves Bilateral Spatial Filtering if the parameter d is greater than zero, which uses a joint spatial and radiometric weighting to keep intensity discontinuities at object boundaries while suppressing noise in smooth areas <ref:2604.22093#pg3>.

Conclusion: Jane: So, looking at the results on the Low Light paired dataset LOL, what did they actually find out about how well FLARE-BO works compared to other methods?

Tom: The paper shows that FLARE-BO achieves the highest PSNR and SSIM of all compared methods by a substantial margin, with its closest competitor on PSNR and SSIM being Rodrigues et al. four <ref:2604.22093#pg1>.

Lu: On NIQE, FLARE-BO scores four point six nine five, which is higher than all other baselines including RetinexNet, ZeroIG, SGZ, LIME seven, Zero-DCE eleven, SCI, EnlightenGAN, and RUAS <ref:2604.22093#pg1>.

Meng: That slightly higher NIQE compared to Rodrigues et al. four is expected because the SSIM weighting in their objective function steers the search away from heavy global smoothing that lowers NIQE but hurts structural detail which SSIM rewards <ref:2604.22093#pg1>.

Lalam: Qualitatively, FLARE-BO produces images with consistent brightness across the full frame, neutral colour rendition, and well preserved structural detail. The illumination normalisation stage is particularly effective in scenes with mixed or spatially varying lighting <ref:2604.22093#pg1>.

Tom: The histogram analysis supports this by showing that FLARE-BO shifts the distribution toward the mid intensity range while maintaining a similar overall shape to the ground truth.

Jane: So, to wrap up, FLARE-BO is a training free lowlight image enhancement framework that improves PSNR by three point six one dB and SSIM by zero point one three two on LOL relative to baseline four.

Lu: Its primary limitation is that it has a higher NIQE, which the authors state is a deliberate consequence of heavily weighting SSIM in the objective function and also the optimiser can only be run offline.

Meng: I have to point out that this optimisation process is slower than deep learning methods, and you can only run this optimiser offline, so it might not be practical for real-time deployment on a fast robot.

Lalam: That's a fair point about the speed; it also means that because it has a higher NIQE, future work will likely need to incorporate perceptual loss terms to recover NIQE competitiveness.

Tom: So, we're looking at FLARE-BO, an extended framework from Rodrigues et al. four, which adds eight parameters and LIME illumination normalisation to handle low light without training data <ref:2604.22093#pg1>.

Jane: It’s a complex system that offers high structural detail and consistent colour rendition for robotic vision in tricky lighting conditions, even if the optimisation process itself isn't as fast as some deep learning alternatives.

Nathan Shankar, Pawel Ladosz, Hujun Yin

University of Manchester

cs.CV, eess.IV

Submitted: 2026-04-23

Updated: 2026-04-23

Comments: 7 pages, 2 tables and 4 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 76/100

The gist: The gist The proposed FLARE-BO framework extends Bayesian optimisation to jointly optimise eight parameters spanning across gamma correction, LIME-style illumination normalisation, chrominance

Key concepts

Bayesian Optimisation Engine: This engine intelligently searches for the best combination of eight image enhancement parameters. It uses techniques like Sobol initialisation and Log Expected Improvement to explore a vast parameter space efficiently. It normalizes parameters to ensure all dimensions are explored equally, leading to principled tuning of the enhancement pipeline.
Structured Eight-Stage Pipeline: The method processes each low-light image through a specific sequence of eight enhancement steps. These stages include illumination correction, gamma adjustment, white balance, chrominance denoising, spatial filtering (bilateral), luminance denoising (NLM), and adaptive post-smoothing. This structured approach ensures comprehensive noise reduction and quality restoration.
Composite Objective Function: The optimization process is guided by a function that prioritizes structural similarity over simple pixel accuracy. It heavily weights Structural Similarity Index Measure (SSIM) at 80% to ensure the output looks perceptually coherent and retains important details, while also including PSNR and NIQE to balance fidelity and perceptual quality for robotic perception.

Terminology

Core — 1× in this paper · explained in 2 episodes

Summary

The gist The proposed FLARE-BO framework extends Bayesian optimisation to jointly optimise eight parameters spanning across gamma correction, LIME-style illumination normalisation, chrominance denoising, bilateral filtering, NLM denoising, GreyWorld automatic white balance, and adaptive post smoothing for low-light image enhancement. This method achieves marked improvements over existing methods when benchmarked using the Low Light paired dataset (LOL)

How it works

FLARE-BO processes each low-light image I through a structured eight-stage pipeline whose parameters θ = (α, β, γ, h, σs, λ, d, hc) are selected per image by Bayesian Optimisation The key contributions include a structured eight-stage enhancement pipeline incorporating LIME style guided illumination normalisation and other specific components controlled by dedicated optimised parameters.

The eight stages of the pipeline are as follows:

Stage 1 involves LIME-Style Illumination Normalisation, where when λ > 0, a Retinex-inspired illumination correction is applied first to estimate the per-pixel illumination map and then normalise the image.
Stage 2 applies Gamma and Linear Contrast Adjustment using the formula I′ = clipα · Iγlime · 255 + β, 0, 255, where α is the linear contrast factor, γ is the gamma exponent, and β is the additive brightness offset.
Stage 3 involves Grey World Automatic White Balance (AWB) correction to remove chromatic bias by assuming that the mean of each colour channel should equal the global mean intensity µ¯ under neutral illumination.
Stage 4 applies LAB Space Chrominance Denoising when hc > 0, converting I′ to the CIELAB colour space and applying a median filter separately to the a∗ and b∗ chrominance channels.
Stage 5 involves Bilateral Spatial Filtering when d > 0, which uses a joint spatial and radiometric weighting to preserve intensity discontinuities at object boundaries while suppressing noise in smooth regions.
Stage 6 applies NLM Luminance Denoising with strength h after bilateral filtering to suppress the residual fine grained luminance noise that bilateral filtering may miss in high-frequency texture regions.
Stage 7 involves Adaptive Gaussian Post-Smoothing with standard deviation σs to suppress quantisation artefacts, where this parameter allows the optimiser to explicitly control the perceptual naturalness of the final output.

Bayesian Optimisation Engine

The search engine employs a unit hypercube parameter normalisation, objective standardisation, Sobol quasi-random initialisation, and Log Expected Improvement acquisition for principled exploration of the expanded space. All raw parameters are mapped to the unit hypercube [0, 1]8 before any GP operation to remove large scale disparity between parameters like α and h. Objective standardisation is performed by transforming collected objective values using ˜fi = fi − ¯f / σf + ϵ, which stabilizes GP kernel hyperparameter optimisation across the wide range of absolute objective values.

The scalar objective function f(θ) prioritises structural similarity with the structure prioritizing composite objective function f(θ) that strongly emphasises structural similarity, steering the search toward perceptually coherent outputs suited to downstream robotic perception. This objective is defined as f(θ) = 1.0 PSNR + 80.0 SSIM − 5.0 NIQE, where the coefficients are set empirically to compensate for the differing numerical scales of each metric.

Experimental Results

Performance of the proposed method is benchmarked using the Low Light paired dataset (LOL) and results show marked improvements of the proposed method over existing methods that were not specifically trained using this dataset. FLARE-BO achieves the highest PSNR and SSIM of all compared methods by a substantial margin, with its closest competitor on PSNR and SSIM being Rodrigues et al. [4].

On NIQE, FLARE-BO scores 4.695, which is higher than all other baselines including RetinexNet, ZeroIG, SGZ, LIME, Zero-DCE, SCI, EnlightenGAN, and RUAS. This moderately higher NIQE relative to [4] is the expected consequence of the SSIM dominant weighting in Equation (10) as the optimiser is steered away from the heavy global smoothing that lowers NIQE but undermines the structural detail that SSIM rewards.

Qualitatively, FLARE-BO produces images with consistent brightness across the full frame, neutral colour rendition, and well preserved structural detail. The illumination normalisation stage is particularly effective in scenes with mixed or spatially varying lighting, where the input contains both heavily underexposed regions and locally brighter objects. The histogram analysis further supports this behaviour by showing that FLARE-BO shifts the distribution toward the mid intensity range while maintaining a similar overall shape to the ground truth.

Conclusion

This paper presented FLARE-BO, a training free lowlight image enhancement framework that extends the Bayesian optimisation of Rodrigues et al. [4]. By expanding from three to eight jointly optimised parameters and adding LIME-style illumination normalisation, gamma correction, Grey-World white balance, LAB-space chrominance denoising, bilateral filtering, and adaptive post-smoothing, FLARE-BO achieves PSNR improvement of +3.61 dB and SSIM improvement of +0.132 on the LOL dataset relative to the baseline [4]. The method requires no training data making it usable on any kind of low-light image without prior scene information. Its primary limitation is that it has a higher NIQE, which is the deliberate consequence of heavily weighting SSIM in the objective function and also the optimisation process is slower than deep learning methods and this optimiser can only be run offline. Future work will explore incorporating perceptual loss terms to recover NIQE competitiveness.

ACKNOWLEDGMENT

This work was supported by the Robotics and Artificial Intelligence Collaboration (RAICo)

REFERENCES

[1] H. Wu, Y. Hou, W. Xu, and M. Zhao, “Ultra-low-lightlevel digital still camera for autonomous underwater vehicle,” Optical Engineering, vol. 58, no. 1, pp. 013 106–013 106, 2019

[2] S. Schwaiger, L. Muster, G. Novotny, M. Schebek, W. Wober, S. Thalhammer, and C. B ¨ ohm, “Ugv-cbrn: an ¨ unmanned ground vehicle for chemical, biological, radiological, and nuclear disaster response,” arXiv preprint arXiv:2406.14385, 2024

[3] M. Abdeh, F. Abut, and F. Akay, “Autonomous navigation in search and rescue simulated environment using deep reinforcement learning,” Balkan Journal of Electrical and Computer Engineering, vol. 9, no. 2, pp. 92–98, 2021

[4] G. F. Rodrigues, A. B. Viana, L. A. Martinho, J. M. Calvalcanti, J. L Pio, and F G Oliveira, “Low-light image quality enhancement through bayesian optimization using gaussian processes,” in 2025 Brazilian Conference on Robotics (CROS), vol 1 IEEE, 2025, pp 1–6

[5] C. Wei, W. Wang, W. Yang, and J Liu, “Deep retinex decomposition for low-light enhancement,” arXiv preprint arXiv:1808.04560, 2018

[6] L H Pham, D N.-N Tran, and J W Jeon, “Lowlight image enhancement for autonomous driving systems using driveretinex-net,” in 2020 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia). IEEE, 2020, pp 1–5

[7] X Guo, Y Li, and H Ling, “Lime: Low-light image enhancement via illumination map estimation,” IEEE Transactions on image processing, vol. 26, no. 2, pp. 982–993, 2016

[8] W Kubinger, M Vincze, and M Ayromlou, “The role of gamma correction in colour image processing,” in 9th European signal processing conference (EUSIPCO 1998). IEEE, 1998, pp 1–4

[9] C.-H Lee, J.-L Shih, C.-C Lien, and C.-C Han, “Adaptive multiscale retinex for image contrast enhancement,” in 2013 International Conference on Signal-Image Technology & Internet-Based Systems (SITIS).

Improvements for AI systems

Adaptive parameter selection for image enhancement allows autonomous robotic systems to maintain reliable visual perception across a wide range of illumination conditions without requiring prior training data. This enables robots performing subsea navigation or structural inspection to operate effectively in uncontrolled, low-light environments by dynamically adjusting eight enhancement parameters per image.
Joint optimization of eight parameters including LIME style illumination normalisation and Grey World automatic white balance corrects spatially non uniform darkness and chromatic bias simultaneously. This means the enhanced system can robustly handle scenes with mixed or spatially varying lighting, restoring visibility in both heavily underexposed regions and locally brighter objects without introducing saturation artefacts or clipping.
Incorporation of LAB space chrominance denoising via a median filter on a∗ and b∗ channels eliminates visually disturbing random colour speckle that luminance only denoisers cannot address. This results in outputs that are both perceptually stable and statistically aligned with ground truth, improving structural correspondence with the ground-truth reference image.
The Bayesian optimisation engine uses unit hypercube parameter normalisation and objective standardisation to ensure the GP kernel assigns equitable length scale learning to all eight dimensions of enhancement parameters. This principled exploration of the expanded space guarantees effective search across the wide range of absolute objective values that arise across different images, leading to a structured remapping of intensities that preserves relative luminance relationships.
The composite objective function f(theta) strongly emphasises structural similarity by weighting SSIM at 80.0 and NIQE at-5.0, steering the search toward perceptually coherent outputs suited to downstream robotic perception while balancing PSNR for reconstruction fidelity. This steering mechanism ensures that the resulting images are both statistically aligned with ground truth and structurally detailed enough for accurate downstream tasks.